train_classification.py

from os.path import join
from glob import glob

import numpy as np
import matplotlib.pyplot as plt

from batch_generators.batch_generator_classification_nist import BatchGenerator_Classification_NIST
from neural_nets.neural_net_classification import NeuralNet_Classification
from neural_nets.tf_utils import visualize_layers, visualize_embedding
import config

########################################
# Set globals
########################################

DATAPATH = join(config.datadir, 'sd04/png_txt')
# META_FILE = 'GeneralPatterns.txt'#'CLASSIFICATION-extended pattern set.pet'
HEIGHT = 512
WIDTH = 512
BATCH_SIZE = 32
NUM_STEPS = 2001
DROPOUT = .5
AUGMENT = 1
DECAY = 1

bg = BatchGenerator_Classification_NIST(path=DATAPATH, height=HEIGHT, width=WIDTH, include_aug=True, n_train=3000)

x, y = bg.generate_train_batch(32)
# print(bg.label_dict)
plt.imshow(x[0].reshape(512, 512), cmap='gray')
plt.show()

nn = NeuralNet_Classification(HEIGHT, WIDTH, len(bg.label_dict))
# nn.load_weights('models/neural_net1000.ckpt')

loss, val_loss = nn.train(num_steps=NUM_STEPS,
                          batchgen=bg,
                          batch_size=BATCH_SIZE,
                          dropout_rate=DROPOUT,
                          augment=AUGMENT,
                          lr=.0001,
                          decay=DECAY)


# plt.plot(loss, color='b', alpha=.7)
# plt.plot(val_loss, color='g', alpha=.7)
# plt.show()

# plt.plot([np.mean(loss[index:index+30]) for index, value in enumerate(loss)], color='b', alpha=.7)
# plt.plot([np.mean(val_loss[index:index+30]) for index, value in enumerate(val_loss)], color='g', alpha=.7)
# plt.show()

########################################
# Determine acc
########################################

def get_acc(bg, train_val):
    samples = 0
    correct = 0
    for i in range(10):
        if train_val == 'train':
            x, y = bg.generate_train_batch(32)
        if train_val == 'val':
            x, y = bg.generate_val_batch(32)

        for img, label in zip(x, y):
            samples += 1
            pred = nn.predict(img)
            if np.argmax(pred) == np.argmax(label):
                correct += 1

    print('{} acc: {}'.format(train_val, correct / samples))


get_acc(bg, 'train')
get_acc(bg, 'val')